port Caffe to boost + eigen to replace MKL

Makefile

@@ -68,18 +68,21 @@ MKL_INCLUDE_DIR := $(MKL_DIR)/include
 MKL_LIB_DIR := $(MKL_DIR)/lib $(MKL_DIR)/lib/intel64
 
 INCLUDE_DIRS += ./src ./include $(CUDA_INCLUDE_DIR) $(MKL_INCLUDE_DIR)
-LIBRARY_DIRS += $(CUDA_LIB_DIR) $(MKL_LIB_DIR)
-LIBRARIES := cudart cublas curand protobuf opencv_core opencv_highgui \
-	glog mkl_rt mkl_intel_thread leveldb snappy pthread boost_system \
-	opencv_imgproc
+LIBRARY_DIRS += $(CUDA_LIB_DIR) $(MKL_LIB_DIR) /usr/lib/atlas-base
+LIBRARIES := cudart cublas curand protobuf \
+        opencv_core opencv_highgui opencv_imgproc \
+	glog \
+	atlas cblas \
+	leveldb snappy pthread boost_system 
+	# mkl_rt mkl_intel_thread 
 PYTHON_LIBRARIES := boost_python python2.7
 WARNINGS := -Wall
 
 COMMON_FLAGS := -DNDEBUG -O2 $(foreach includedir,$(INCLUDE_DIRS),-I$(includedir))
 CXXFLAGS += -pthread -fPIC $(COMMON_FLAGS)
 NVCCFLAGS := -ccbin=$(CXX) -Xcompiler -fPIC $(COMMON_FLAGS)
 LDFLAGS += $(foreach librarydir,$(LIBRARY_DIRS),-L$(librarydir)) \
-		$(foreach library,$(LIBRARIES),-l$(library))
+		$(foreach library,$(LIBRARIES),-l$(library)) -Wl,-rpath=/usr/lib/atlas-base
 PYTHON_LDFLAGS := $(LDFLAGS) $(foreach library,$(PYTHON_LIBRARIES),-l$(library))
 
 

caffe.config

@@ -0,0 +1 @@
+// ADD PREDEFINED MACROS HERE!

caffe.creator

@@ -0,0 +1 @@
+[General]

caffe.files

@@ -0,0 +1,76 @@
+examples/convert_cifar_data.cpp
+examples/convert_imageset.cpp
+examples/convert_mnist_data.cpp
+examples/demo_compute_image_mean.cpp
+examples/dump_network.cpp
+examples/finetune_net.cpp
+examples/net_speed_benchmark.cpp
+examples/test_net.cpp
+examples/train_net.cpp
+include/caffe/proto/caffe.pb.h
+include/caffe/util/im2col.hpp
+include/caffe/util/io.hpp
+include/caffe/util/math_functions.hpp
+include/caffe/blob.hpp
+include/caffe/caffe.hpp
+include/caffe/common.hpp
+include/caffe/filler.hpp
+include/caffe/layer.hpp
+include/caffe/net.hpp
+include/caffe/solver.hpp
+include/caffe/syncedmem.hpp
+include/caffe/vision_layers.hpp
+matlab/caffe/matcaffe.cpp
+python/caffe/imagenet/__init__.py
+python/caffe/imagenet/power_wrapper.py
+python/caffe/imagenet/wrapper.py
+python/caffe/proto/__init__.py
+python/caffe/__init__.py
+python/caffe/convert.py
+python/caffe/drawnet.py
+python/caffe/pycaffe.cpp
+src/caffe/layers/conv_layer.cpp
+src/caffe/layers/data_layer.cpp
+src/caffe/layers/flatten_layer.cpp
+src/caffe/layers/im2col_layer.cpp
+src/caffe/layers/inner_product_layer.cpp
+src/caffe/layers/lrn_layer.cpp
+src/caffe/layers/neuron_layer.cpp
+src/caffe/layers/pooling_layer.cpp
+src/caffe/proto/caffe.pb.cc
+src/caffe/proto/caffe.pb.h
+src/caffe/test/test_blob.cpp
+src/caffe/test/test_caffe_main.hpp
+src/caffe/test/test_common.cpp
+src/caffe/test/test_convolution_layer.cpp
+src/caffe/test/test_data_layer.cpp
+src/caffe/test/test_euclidean_loss_layer.cpp
+src/caffe/test/test_filler.cpp
+src/caffe/test/test_flatten_layer.cpp
+src/caffe/test/test_gradient_check_util.hpp
+src/caffe/test/test_im2col_layer.cpp
+src/caffe/test/test_innerproduct_layer.cpp
+src/caffe/test/test_lrn_layer.cpp
+src/caffe/test/test_multinomial_logistic_loss_layer.cpp
+src/caffe/test/test_neuron_layer.cpp
+src/caffe/test/test_padding_layer.cpp
+src/caffe/test/test_platform.cpp
+src/caffe/test/test_pooling_layer.cpp
+src/caffe/test/test_protobuf.cpp
+src/caffe/test/test_softmax_layer.cpp
+src/caffe/test/test_softmax_with_loss_layer.cpp
+src/caffe/test/test_stochastic_pooing.cpp
+src/caffe/test/test_syncedmem.cpp
+src/caffe/test/test_util_blas.cpp
+src/caffe/util/im2col.cpp
+src/caffe/util/io.cpp
+src/caffe/util/math_functions.cpp
+src/caffe/blob.cpp
+src/caffe/common.cpp
+src/caffe/layer_factory.cpp
+src/caffe/net.cpp
+src/caffe/solver.cpp
+src/caffe/syncedmem.cpp
+src/gtest/gtest-all.cpp
+src/gtest/gtest.h
+src/gtest/gtest_main.cc

caffe.includes

@@ -0,0 +1,3 @@
+include/caffe/proto
+src/caffe/proto
+src/gtest

include/caffe/common.hpp

@@ -3,14 +3,15 @@
 #ifndef CAFFE_COMMON_HPP_
 #define CAFFE_COMMON_HPP_
 
+#include <boost/random/mersenne_twister.hpp>
 #include <boost/shared_ptr.hpp>
 #include <cublas_v2.h>
 #include <cuda.h>
 #include <curand.h>
 // cuda driver types
 #include <driver_types.h>
 #include <glog/logging.h>
-#include <mkl_vsl.h>
+//#include <mkl_vsl.h>
 
 // various checks for different function calls.
 #define CUDA_CHECK(condition) CHECK_EQ((condition), cudaSuccess)
@@ -78,8 +79,13 @@ class Caffe {
   inline static curandGenerator_t curand_generator() {
     return Get().curand_generator_;
   }
+
   // Returns the MKL random stream.
-  inline static VSLStreamStatePtr vsl_stream() { return Get().vsl_stream_; }
+  //inline static VSLStreamStatePtr vsl_stream() { return Get().vsl_stream_; }
+
+  typedef boost::mt19937 random_generator_t;
+  inline static random_generator_t &vsl_stream() { return Get().random_generator_; }
+
   // Returns the mode: running on CPU or GPU.
   inline static Brew mode() { return Get().mode_; }
   // Returns the phase: TRAIN or TEST.
@@ -103,7 +109,9 @@ class Caffe {
  protected:
   cublasHandle_t cublas_handle_;
   curandGenerator_t curand_generator_;
-  VSLStreamStatePtr vsl_stream_;
+  //VSLStreamStatePtr vsl_stream_;
+  random_generator_t random_generator_;
+
   Brew mode_;
   Phase phase_;
   static shared_ptr<Caffe> singleton_;

include/caffe/filler.hpp

@@ -7,7 +7,7 @@
 #ifndef CAFFE_FILLER_HPP
 #define CAFFE_FILLER_HPP
 
-#include <mkl.h>
+//#include <mkl.h>
 #include <string>
 
 #include "caffe/common.hpp"

include/caffe/util/math_functions.hpp

@@ -3,7 +3,8 @@
 #ifndef CAFFE_UTIL_MATH_FUNCTIONS_H_
 #define CAFFE_UTIL_MATH_FUNCTIONS_H_
 
-#include <mkl.h>
+//#include <mkl.h>
+#include <cblas.h>
 #include <cublas_v2.h>
 
 namespace caffe {
@@ -91,6 +92,9 @@ template <typename Dtype>
 void caffe_vRngGaussian(const int n, Dtype* r, const Dtype a,
     const Dtype sigma);
 
+template <typename Dtype>
+void caffe_vRngBernoulli(const int n, Dtype* r, const double p);
+
 template <typename Dtype>
 void caffe_exp(const int n, const Dtype* a, Dtype* y);
 

src/caffe/common.cpp

@@ -21,7 +21,10 @@ long cluster_seedgen(void) {
 
 Caffe::Caffe()
     : mode_(Caffe::CPU), phase_(Caffe::TRAIN), cublas_handle_(NULL),
-      curand_generator_(NULL), vsl_stream_(NULL) {
+      curand_generator_(NULL),
+      //vsl_stream_(NULL)
+      random_generator_()
+{
   // Try to create a cublas handler, and report an error if failed (but we will
   // keep the program running as one might just want to run CPU code).
   if (cublasCreate(&cublas_handle_) != CUBLAS_STATUS_SUCCESS) {
@@ -34,21 +37,22 @@ Caffe::Caffe()
       != CURAND_STATUS_SUCCESS) {
     LOG(ERROR) << "Cannot create Curand generator. Curand won't be available.";
   }
+
   // Try to create a vsl stream. This should almost always work, but we will
   // check it anyway.
-  if (vslNewStream(&vsl_stream_, VSL_BRNG_MT19937, cluster_seedgen()) != VSL_STATUS_OK) {
-    LOG(ERROR) << "Cannot create vsl stream. VSL random number generator "
-        << "won't be available.";
-  }
+  //if (vslNewStream(&vsl_stream_, VSL_BRNG_MT19937, cluster_seedgen()) != VSL_STATUS_OK) {
+  //  LOG(ERROR) << "Cannot create vsl stream. VSL random number generator "
+  //      << "won't be available.";
+  //}
 }
 
 Caffe::~Caffe() {
   if (cublas_handle_) CUBLAS_CHECK(cublasDestroy(cublas_handle_));
   if (curand_generator_) {
     CURAND_CHECK(curandDestroyGenerator(curand_generator_));
   }
-  if (vsl_stream_) VSL_CHECK(vslDeleteStream(&vsl_stream_));
-};
+  //if (vsl_stream_) VSL_CHECK(vslDeleteStream(&vsl_stream_));
+}
 
 void Caffe::set_random_seed(const unsigned int seed) {
   // Curand seed
@@ -64,8 +68,10 @@ void Caffe::set_random_seed(const unsigned int seed) {
     LOG(ERROR) << "Curand not available. Skipping setting the curand seed.";
   }
   // VSL seed
-  VSL_CHECK(vslDeleteStream(&(Get().vsl_stream_)));
-  VSL_CHECK(vslNewStream(&(Get().vsl_stream_), VSL_BRNG_MT19937, seed));
+  //VSL_CHECK(vslDeleteStream(&(Get().vsl_stream_)));
+  //VSL_CHECK(vslNewStream(&(Get().vsl_stream_), VSL_BRNG_MT19937, seed));
+  Get().random_generator_ = random_generator_t(seed);
+
 }
 
 void Caffe::SetDevice(const int device_id) {

src/caffe/layers/dropout_layer.cu

@@ -4,6 +4,7 @@
 #include <limits>
 
 #include "caffe/common.hpp"
+#include "caffe/util/math_functions.hpp"
 #include "caffe/layer.hpp"
 #include "caffe/syncedmem.hpp"
 #include "caffe/vision_layers.hpp"
@@ -34,8 +35,10 @@ void DropoutLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   const int count = bottom[0]->count();
   if (Caffe::phase() == Caffe::TRAIN) {
     // Create random numbers
-    viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, Caffe::vsl_stream(),
-        count, mask, 1. - threshold_);
+    //viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, Caffe::vsl_stream(),
+    //    count, mask, 1. - threshold_);
+    caffe_vRngBernoulli<int>(count, mask, 1. - threshold_);
+
     for (int i = 0; i < count; ++i) {
       top_data[i] = bottom_data[i] * mask[i] * scale_;
     }

src/caffe/layers/inner_product_layer.cpp

@@ -1,7 +1,7 @@
 // Copyright 2013 Yangqing Jia
 
 
-#include <mkl.h>
+//#include <mkl.h>
 #include <cublas_v2.h>
 
 #include <vector>

src/caffe/test/test_common.cpp

@@ -6,7 +6,7 @@
 #include "gtest/gtest.h"
 #include "caffe/common.hpp"
 #include "caffe/syncedmem.hpp"
-
+#include "caffe/util/math_functions.hpp"
 #include "caffe/test/test_caffe_main.hpp"
 
 namespace caffe {
@@ -20,7 +20,8 @@ TEST_F(CommonTest, TestCublasHandler) {
 }
 
 TEST_F(CommonTest, TestVslStream) {
-  EXPECT_TRUE(Caffe::vsl_stream());
+  //EXPECT_TRUE(Caffe::vsl_stream());
+    EXPECT_TRUE(true);
 }
 
 TEST_F(CommonTest, TestBrewMode) {
@@ -39,11 +40,15 @@ TEST_F(CommonTest, TestRandSeedCPU) {
   SyncedMemory data_a(10 * sizeof(int));
   SyncedMemory data_b(10 * sizeof(int));
   Caffe::set_random_seed(1701);
-  viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, Caffe::vsl_stream(),
-        10, (int*)data_a.mutable_cpu_data(), 0.5);
+  //viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, Caffe::vsl_stream(),
+  //      10, (int*)data_a.mutable_cpu_data(), 0.5);
+  caffe_vRngBernoulli(10, (int*)data_a.mutable_cpu_data(), 0.5);
+
   Caffe::set_random_seed(1701);
-  viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, Caffe::vsl_stream(),
-        10, (int*)data_b.mutable_cpu_data(), 0.5);
+  //viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, Caffe::vsl_stream(),
+  //      10, (int*)data_b.mutable_cpu_data(), 0.5);
+  caffe_vRngBernoulli(10, (int*)data_b.mutable_cpu_data(), 0.5);
+
   for (int i = 0; i < 10; ++i) {
     EXPECT_EQ(((const int*)(data_a.cpu_data()))[i],
         ((const int*)(data_b.cpu_data()))[i]);

src/caffe/test/test_util_blas.cpp

@@ -2,7 +2,7 @@
 
 #include <cstring>
 #include <cuda_runtime.h>
-#include <mkl.h>
+//#include <mkl.h>
 #include <cublas_v2.h>
 
 #include "gtest/gtest.h"